roach-strategy（蟑螂策略）

🪳 ROACH v2.0 — Striker Only. v2-Runtime-Native. Maker Exits.

Cockroaches survive anything. ROACH survives by not trading when there's no explosion.

What changed in v2.0

Architecture (not thesis):

Why v2: v1 used MARKET orders for every exit, paying ~3 bp/exit in HL taker fees. Recent 25-trade sample: $46 in HL fees, 100% taker. v2's FEE_OPTIMIZED_LIMIT with ensure_execution_as_taker: true and 60s timeout prefers maker fill but accepts taker fallback when the position must close. Expected recovery: 50-70% of HL exit fees.

Thesis preserved from v1.2: Stalker disabled. Striker only. FIRST_JUMP from #25+, rank jump >= 15 OR contribVelocity >= 15, score >= 10 with 4+ reasons, 1.5x volume, 1h price aligned, 4h trend aligned, XYZ banned.

Files

Producer behavior

Runs every 90s via cron. On each tick:

1. Reentrancy guard: acquires state/<wallet-hash>/producer.lock. If a prior run hasn't released it (cron faster than MCP latency), this run skips cleanly.
2. Fetch markets: leaderboard_get_markets (top 50, XYZ filtered out at parse time).
3. Detect striker signals: rank jumps, contribVelocity, contribution explosions vs scan history.
4. Apply hard gates: 4h trend aligned, cc_15m >= 0.5, 1h price aligned >= 0.1%, volume >= 1.5x, score >= 10, 4+ reasons, asset not in 120min cooldown.
5. Emit signals via openclaw senpi external-scanner ingest.
6. Persist scan history + cooldown state under state/<wallet-hash>/.

NO execution code. NO position-tracking. NO DSL state. The runtime owns all of that.

Entry flow

Producer cron (90s)
  ↓ Striker signal detected
  ↓ external-scanner ingest --scanner roach_signals
Runtime
  ↓ Schema-validates fields against runtime.yaml
  ↓ LLM gate (decision_model = ${ROACH_DECISION_MODEL})
  ↓ Pass-through unless malformed (producer already filtered everything else)
  ↓ OPEN_POSITION via FEE_OPTIMIZED_LIMIT (maker-first, 60s, taker fallback)
DSL (runtime-managed)
  ↓ Phase 1 max_loss 18% / 3-breach
  ↓ Phase 2 tiers (7/40, 12/55, 15/75, 20/85)
  ↓ hard_timeout 120min, weak_peak 45min, dead_weight 25min
  ↓ Exit via FEE_OPTIMIZED_LIMIT (maker-first, 60s, taker fallback)

Required env vars

The runtime YAML uses these substitutions:

The producer reads:

Producer install (on OpenClaw host)

Source path in the senpi-skills repo: roach/. Install destination on the OpenClaw host: /data/workspace/skills/roach-strategy/. The two names differ — repo uses roach, install dir uses the SKILL.md name field (roach-strategy).

# 1. Pull the skill files (source: senpi-skills/main/roach/)
mkdir -p /data/workspace/skills/roach-strategy/scripts
mkdir -p /data/workspace/skills/roach-strategy/config

curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/roach/runtime.yaml \
  -o /data/workspace/skills/roach-strategy/runtime.yaml
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/roach/scripts/roach-producer.py \
  -o /data/workspace/skills/roach-strategy/scripts/roach-producer.py
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/roach/scripts/roach_config.py \
  -o /data/workspace/skills/roach-strategy/scripts/roach_config.py
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/roach/config/roach-config.json \
  -o /data/workspace/skills/roach-strategy/config/roach-config.json

# Remove the v1 scanner if it's still there from a prior install
rm -f /data/workspace/skills/roach-strategy/scripts/roach-scanner.py

# 2. Install the runtime
# ROACH_DECISION_MODEL is REQUIRED — pick any model supported by the runtime's
# model registry. Examples: gemini-3.1-pro-preview, claude-sonnet-4-20250514.
# Pass BARE model name only — NO provider prefix (OpenClaw double-prefixes).
WALLET_ADDRESS=0x... \
TELEGRAM_CHAT_ID=... \
ROACH_DECISION_MODEL=gemini-3.1-pro-preview \
  openclaw senpi runtime create --path /data/workspace/skills/roach-strategy/runtime.yaml

# 3. Schedule the producer (~90s cadence)
# ROACH_WALLET (NOT generic STRATEGY_ADDRESS) is required by the producer.
# Producer fails loud at startup if not set — misconfig is visible immediately.
#
# RECOMMENDED PATTERN: cron-as-Claude-turn with NO_REPLY filter.
# Each cron tick spins up an isolated session, runs the producer, and
# replies to TG ONLY if there's a non-zero event (signal pushed, error,
# unexpected state). On the silence-is-correct ticks (most of them for
# Roach), the agent stays quiet — zero TG noise, zero log file growth.
# Cron run records are still accessible via `openclaw cron runs --id <id>`.
openclaw cron add \
  --name "roach-v2-producer" \
  --cron "*/2 * * * *" \
  --session isolated \
  --wake now \
  --message $'export ROACH_WALLET=0x... && python3 /data/workspace/skills/roach-strategy/scripts/roach-producer.py\n\nCRITICAL INSTRUCTION: If the output shows \'candidates\': 0 and \'status\': \'ok\', you MUST reply EXACTLY with NO_REPLY to remain silent. Only report to the user if there is an error, a crash, or if signals are actually pushed.' \
  --no-deliver

# Note: openclaw cron resolution is 1 minute. The producer is designed for
# 90s cadence; */2 (every 2 min) is the closest cron-supported approximation.
# Some installs run faster via the underlying scheduler (~90s observed).
# The fcntl reentrancy guard handles any overlap safely.

# Alternative pattern: shell-redirect (cheaper, no Claude tokens per tick,
# but generates a log file that needs rotation). Use this if you prefer
# to grep raw producer JSON output:
#
#   --message "Run \`SENPI_API_KEY=<KEY> ROACH_WALLET=0x... python3 /data/workspace/skills/roach-strategy/scripts/roach-producer.py >> /var/log/openclaw/roach-v2.log 2>&1\` and report success/failure in this log."

# 4. Verify
openclaw senpi runtime list                        # expect: roach-tracker v1.3.0
openclaw senpi status --runtime roach-tracker

# Cron-as-Claude-turn pattern — verify via OpenClaw's structured run log:
openclaw cron list --json                          # find the cron id
openclaw cron runs --id <cron-id> --limit 5        # last 5 tick records (ts, status, durationMs, nextRunAtMs)

# Inspect producer state (this is the same regardless of cron pattern):
ls -la /data/workspace/skills/roach-strategy/state/<wallet-hash>/
# Expect: producer.lock + scan-history.json with mtime in last few minutes.
# A fresh scan-history.json (mtime < 5 min) confirms the cron is firing.

Risk envelope (declarative, runtime-enforced)

Expected behavior

SILENCE IS CORRECT. If Roach goes 24-48 hours without a trade, that means there were no FIRST_JUMP explosions worth taking. That's the experiment working, not a bug.

What was removed in v2

Removed (responsibilities moved to runtime):

• roach-scanner.py (replaced by roach-producer.py)
• Trade counter / daily-loss tracking (runtime guard_rails)
• Bootstrap gate / config/bootstrap-complete.json (runtime install handles state)
• Asset cooldown bookkeeping in scanner (still in producer as defense-in-depth, but runtime is the authority)
• Dynamic slots based on PnL (runtime owns slots)
• All create_position calls (runtime owns execution)

Permanently disabled (decided in v1.0):

• Stalker mode and all accumulation detection
• XYZ equities (banned at scan level)

License

MIT — Built by Senpi (https://senpi.ai). Source: https://github.com/Senpi-ai/senpi-skills

Skill Attribution

When creating a strategy, include skill_name and skill_version in the call. See references/skill-attribution.md for details.